v0.5.0.

Release Notes (Post v0.2.0 → v0.5.0)

What’s New

• Hardware Sovereignty (Substrate Equilibrium v0.5.0)
  • Enforced VRAM ceiling at 9,750 MB (~79% of 12 GB) to prevent DWM/driver contention.
  • Embeddings locked to NPU for PCIe offload; status via manage.py substrate.
  • Default local run supports --substrate-equilibrium to auto-apply safe limits.
• Cognitive Safety Gate
  • ImmuneValidator integrated as mandatory middleware in the cognitive pipeline.
• Skill Expansion
  • Added model search/orchestration skill src/skills/integration/demodelis_ganesha (underscore path is canonical).
• Benchmarks & Validation
  • benchmarks/collision_test.py + docs/substrate_collision_test.md to measure LLM TPS under embedding load (contention vs equilibrium).
  • Expected: <10% TPS variance during equilibrium runs.
• Docs & Licensing
  • Clarified dual licensing: Apache 2.0 + Cloudhabil Skills Additional License (CSAL).
  • README/CHANGELOG and eval docs updated for current structure and safety posture.
• Maintenance & Security
  • Dependency bumps across pip/npm (requests, jinja2, js-yaml, vite, etc.).
  • Hygiene fixes in model routing/runtime to respect new resource ceilings.

Upgrade Notes

• Preferred local run: python manage.py local --substrate-equilibrium.
• Verify hardware routing: python manage.py substrate.
• Run collision test in both modes and confirm TPS drop <10% under equilibrium:
  • python benchmarks/collision_test.py --embedding-count 1000 --duration 30 --embedding-delay 5 --model gpia_core.

Compliance & Auditability

• No manifold HTML scaffold included (prior scaffold reverted).
• Skills: underscore path is canonical; hyphenated demodelis path not used.
• ImmuneValidator enforces pre-execution checks across the cognitive pipeline.

Scope

• Changes listed are post-v0.2.0 and present on main.
• No API-breaking changes; reinstall/refresh dependencies recommended.
• Data/config files are unchanged by this release tag; hardware limits are applied at runtime via CLI/env.